1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright 2006 Jake Moilanen <moilanen@austin.ibm.com>, IBM Corp.
4 * Copyright 2006-2007 Michael Ellerman, IBM Corp.
7 #include <linux/crash_dump.h>
8 #include <linux/device.h>
10 #include <linux/irqdomain.h>
11 #include <linux/msi.h>
14 #include <asm/hw_irq.h>
15 #include <asm/ppc-pci.h>
16 #include <asm/machdep.h>
21 static int query_token, change_token;
23 #define RTAS_QUERY_FN 0
24 #define RTAS_CHANGE_FN 1
25 #define RTAS_RESET_FN 2
26 #define RTAS_CHANGE_MSI_FN 3
27 #define RTAS_CHANGE_MSIX_FN 4
28 #define RTAS_CHANGE_32MSI_FN 5
32 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
34 u32 addr, seq_num, rtas_ret[3];
38 addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
39 buid = pdn->phb->buid;
43 if (func == RTAS_CHANGE_MSI_FN || func == RTAS_CHANGE_MSIX_FN ||
44 func == RTAS_CHANGE_32MSI_FN)
45 rc = rtas_call(change_token, 6, 4, rtas_ret, addr,
46 BUID_HI(buid), BUID_LO(buid),
47 func, num_irqs, seq_num);
49 rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
50 BUID_HI(buid), BUID_LO(buid),
51 func, num_irqs, seq_num);
53 seq_num = rtas_ret[1];
54 } while (rtas_busy_delay(rc));
57 * If the RTAS call succeeded, return the number of irqs allocated.
58 * If not, make sure we return a negative error code.
65 pr_debug("rtas_msi: ibm,change_msi(func=%d,num=%d), got %d rc = %d\n",
66 func, num_irqs, rtas_ret[0], rc);
71 static void rtas_disable_msi(struct pci_dev *pdev)
75 pdn = pci_get_pdn(pdev);
80 * disabling MSI with the explicit interface also disables MSI-X
82 if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
84 * may have failed because explicit interface is not
87 if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
88 pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
93 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
95 u32 addr, rtas_ret[2];
99 addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
100 buid = pdn->phb->buid;
103 rc = rtas_call(query_token, 4, 3, rtas_ret, addr,
104 BUID_HI(buid), BUID_LO(buid), offset);
105 } while (rtas_busy_delay(rc));
108 pr_debug("rtas_msi: error (%d) querying source number\n", rc);
115 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
117 struct device_node *dn;
121 dn = pci_device_to_OF_node(pdev);
123 p = of_get_property(dn, prop_name, NULL);
125 pr_debug("rtas_msi: No %s on %pOF\n", prop_name, dn);
129 req_msi = be32_to_cpup(p);
130 if (req_msi < nvec) {
131 pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
133 if (req_msi == 0) /* Be paranoid */
142 static int check_req_msi(struct pci_dev *pdev, int nvec)
144 return check_req(pdev, nvec, "ibm,req#msi");
147 static int check_req_msix(struct pci_dev *pdev, int nvec)
149 return check_req(pdev, nvec, "ibm,req#msi-x");
152 /* Quota calculation */
154 static struct device_node *__find_pe_total_msi(struct device_node *node, int *total)
156 struct device_node *dn;
159 dn = of_node_get(node);
161 p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
163 pr_debug("rtas_msi: found prop on dn %pOF\n",
165 *total = be32_to_cpup(p);
169 dn = of_get_next_parent(dn);
175 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
177 return __find_pe_total_msi(pci_device_to_OF_node(dev), total);
180 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
182 struct device_node *dn;
183 struct eeh_dev *edev;
185 /* Found our PE and assume 8 at that point. */
187 dn = pci_device_to_OF_node(dev);
191 /* Get the top level device in the PE */
192 edev = pdn_to_eeh_dev(PCI_DN(dn));
194 edev = list_first_entry(&edev->pe->edevs, struct eeh_dev,
196 dn = pci_device_to_OF_node(edev->pdev);
200 /* We actually want the parent */
201 dn = of_get_parent(dn);
205 /* Hardcode of 8 for old firmwares */
207 pr_debug("rtas_msi: using PE dn %pOF\n", dn);
213 struct device_node *requestor;
221 static void *count_non_bridge_devices(struct device_node *dn, void *data)
223 struct msi_counts *counts = data;
227 pr_debug("rtas_msi: counting %pOF\n", dn);
229 p = of_get_property(dn, "class-code", NULL);
230 class = p ? be32_to_cpup(p) : 0;
232 if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
233 counts->num_devices++;
238 static void *count_spare_msis(struct device_node *dn, void *data)
240 struct msi_counts *counts = data;
244 if (dn == counts->requestor)
245 req = counts->request;
247 /* We don't know if a driver will try to use MSI or MSI-X,
248 * so we just have to punt and use the larger of the two. */
250 p = of_get_property(dn, "ibm,req#msi", NULL);
252 req = be32_to_cpup(p);
254 p = of_get_property(dn, "ibm,req#msi-x", NULL);
256 req = max(req, (int)be32_to_cpup(p));
259 if (req < counts->quota)
260 counts->spare += counts->quota - req;
261 else if (req > counts->quota)
262 counts->over_quota++;
267 static int msi_quota_for_device(struct pci_dev *dev, int request)
269 struct device_node *pe_dn;
270 struct msi_counts counts;
273 pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
276 pe_dn = find_pe_total_msi(dev, &total);
278 pe_dn = find_pe_dn(dev, &total);
281 pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
285 pr_debug("rtas_msi: found PE %pOF\n", pe_dn);
287 memset(&counts, 0, sizeof(struct msi_counts));
289 /* Work out how many devices we have below this PE */
290 pci_traverse_device_nodes(pe_dn, count_non_bridge_devices, &counts);
292 if (counts.num_devices == 0) {
293 pr_err("rtas_msi: found 0 devices under PE for %s\n",
298 counts.quota = total / counts.num_devices;
299 if (request <= counts.quota)
302 /* else, we have some more calculating to do */
303 counts.requestor = pci_device_to_OF_node(dev);
304 counts.request = request;
305 pci_traverse_device_nodes(pe_dn, count_spare_msis, &counts);
307 /* If the quota isn't an integer multiple of the total, we can
308 * use the remainder as spare MSIs for anyone that wants them. */
309 counts.spare += total % counts.num_devices;
311 /* Divide any spare by the number of over-quota requestors */
312 if (counts.over_quota)
313 counts.quota += counts.spare / counts.over_quota;
315 /* And finally clamp the request to the possibly adjusted quota */
316 request = min(counts.quota, request);
318 pr_debug("rtas_msi: request clamped to quota %d\n", request);
325 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
327 u32 addr_hi, addr_lo;
330 * We should only get in here for IODA1 configs. This is based on the
331 * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
332 * support, and we are in a PCIe Gen2 slot.
335 "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
336 pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
337 addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
338 pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
339 pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
342 static int rtas_prepare_msi_irqs(struct pci_dev *pdev, int nvec_in, int type,
343 msi_alloc_info_t *arg)
348 int use_32bit_msi_hack = 0;
350 if (type == PCI_CAP_ID_MSIX)
351 rc = check_req_msix(pdev, nvec);
353 rc = check_req_msi(pdev, nvec);
358 quota = msi_quota_for_device(pdev, nvec);
360 if (quota && quota < nvec)
364 * Firmware currently refuse any non power of two allocation
365 * so we round up if the quota will allow it.
367 if (type == PCI_CAP_ID_MSIX) {
368 int m = roundup_pow_of_two(nvec);
369 quota = msi_quota_for_device(pdev, m);
375 pdn = pci_get_pdn(pdev);
378 * Try the new more explicit firmware interface, if that fails fall
379 * back to the old interface. The old interface is known to never
383 if (type == PCI_CAP_ID_MSI) {
384 if (pdev->no_64bit_msi) {
385 rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
388 * We only want to run the 32 bit MSI hack below if
389 * the max bus speed is Gen2 speed
391 if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
394 use_32bit_msi_hack = 1;
400 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
403 pr_debug("rtas_msi: trying the old firmware call.\n");
404 rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
407 if (use_32bit_msi_hack && rc > 0)
408 rtas_hack_32bit_msi_gen2(pdev);
410 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
413 if (nvec != nvec_in) {
417 pr_debug("rtas_msi: rtas_change_msi() failed\n");
424 static int pseries_msi_ops_prepare(struct irq_domain *domain, struct device *dev,
425 int nvec, msi_alloc_info_t *arg)
427 struct pci_dev *pdev = to_pci_dev(dev);
428 int type = pdev->msix_enabled ? PCI_CAP_ID_MSIX : PCI_CAP_ID_MSI;
430 return rtas_prepare_msi_irqs(pdev, nvec, type, arg);
434 * ->msi_free() is called before irq_domain_free_irqs_top() when the
435 * handler data is still available. Use that to clear the XIVE
438 static void pseries_msi_ops_msi_free(struct irq_domain *domain,
439 struct msi_domain_info *info,
443 xive_irq_free_data(irq);
447 * RTAS can not disable one MSI at a time. It's all or nothing. Do it
448 * at the end after all IRQs have been freed.
450 static void pseries_msi_post_free(struct irq_domain *domain, struct device *dev)
452 if (WARN_ON_ONCE(!dev_is_pci(dev)))
455 rtas_disable_msi(to_pci_dev(dev));
458 static struct msi_domain_ops pseries_pci_msi_domain_ops = {
459 .msi_prepare = pseries_msi_ops_prepare,
460 .msi_free = pseries_msi_ops_msi_free,
461 .msi_post_free = pseries_msi_post_free,
464 static void pseries_msi_shutdown(struct irq_data *d)
467 if (d->chip->irq_shutdown)
468 d->chip->irq_shutdown(d);
471 static void pseries_msi_mask(struct irq_data *d)
474 irq_chip_mask_parent(d);
477 static void pseries_msi_unmask(struct irq_data *d)
479 pci_msi_unmask_irq(d);
480 irq_chip_unmask_parent(d);
483 static void pseries_msi_write_msg(struct irq_data *data, struct msi_msg *msg)
485 struct msi_desc *entry = irq_data_get_msi_desc(data);
488 * Do not update the MSIx vector table. It's not strictly necessary
489 * because the table is initialized by the underlying hypervisor, PowerVM
490 * or QEMU/KVM. However, if the MSIx vector entry is cleared, any further
491 * activation will fail. This can happen in some drivers (eg. IPR) which
492 * deactivate an IRQ used for testing MSI support.
497 static struct irq_chip pseries_pci_msi_irq_chip = {
498 .name = "pSeries-PCI-MSI",
499 .irq_shutdown = pseries_msi_shutdown,
500 .irq_mask = pseries_msi_mask,
501 .irq_unmask = pseries_msi_unmask,
502 .irq_eoi = irq_chip_eoi_parent,
503 .irq_write_msi_msg = pseries_msi_write_msg,
508 * Set MSI_FLAG_MSIX_CONTIGUOUS as there is no way to express to
509 * firmware to request a discontiguous or non-zero based range of
510 * MSI-X entries. Core code will reject such setup attempts.
512 static struct msi_domain_info pseries_msi_domain_info = {
513 .flags = (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
514 MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX |
515 MSI_FLAG_MSIX_CONTIGUOUS),
516 .ops = &pseries_pci_msi_domain_ops,
517 .chip = &pseries_pci_msi_irq_chip,
520 static void pseries_msi_compose_msg(struct irq_data *data, struct msi_msg *msg)
522 __pci_read_msi_msg(irq_data_get_msi_desc(data), msg);
525 static struct irq_chip pseries_msi_irq_chip = {
526 .name = "pSeries-MSI",
527 .irq_shutdown = pseries_msi_shutdown,
528 .irq_mask = irq_chip_mask_parent,
529 .irq_unmask = irq_chip_unmask_parent,
530 .irq_eoi = irq_chip_eoi_parent,
531 .irq_set_affinity = irq_chip_set_affinity_parent,
532 .irq_compose_msi_msg = pseries_msi_compose_msg,
535 static int pseries_irq_parent_domain_alloc(struct irq_domain *domain, unsigned int virq,
536 irq_hw_number_t hwirq)
538 struct irq_fwspec parent_fwspec;
541 parent_fwspec.fwnode = domain->parent->fwnode;
542 parent_fwspec.param_count = 2;
543 parent_fwspec.param[0] = hwirq;
544 parent_fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
546 ret = irq_domain_alloc_irqs_parent(domain, virq, 1, &parent_fwspec);
553 static int pseries_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
554 unsigned int nr_irqs, void *arg)
556 struct pci_controller *phb = domain->host_data;
557 msi_alloc_info_t *info = arg;
558 struct msi_desc *desc = info->desc;
559 struct pci_dev *pdev = msi_desc_to_pci_dev(desc);
563 hwirq = rtas_query_irq_number(pci_get_pdn(pdev), desc->msi_index);
565 dev_err(&pdev->dev, "Failed to query HW IRQ: %d\n", hwirq);
569 dev_dbg(&pdev->dev, "%s bridge %pOF %d/%x #%d\n", __func__,
570 phb->dn, virq, hwirq, nr_irqs);
572 for (i = 0; i < nr_irqs; i++) {
573 ret = pseries_irq_parent_domain_alloc(domain, virq + i, hwirq + i);
577 irq_domain_set_hwirq_and_chip(domain, virq + i, hwirq + i,
578 &pseries_msi_irq_chip, domain->host_data);
584 /* TODO: handle RTAS cleanup in ->msi_finish() ? */
585 irq_domain_free_irqs_parent(domain, virq, i - 1);
589 static void pseries_irq_domain_free(struct irq_domain *domain, unsigned int virq,
590 unsigned int nr_irqs)
592 struct irq_data *d = irq_domain_get_irq_data(domain, virq);
593 struct pci_controller *phb = irq_data_get_irq_chip_data(d);
595 pr_debug("%s bridge %pOF %d #%d\n", __func__, phb->dn, virq, nr_irqs);
597 /* XIVE domain data is cleared through ->msi_free() */
600 static const struct irq_domain_ops pseries_irq_domain_ops = {
601 .alloc = pseries_irq_domain_alloc,
602 .free = pseries_irq_domain_free,
605 static int __pseries_msi_allocate_domains(struct pci_controller *phb,
608 struct irq_domain *parent = irq_get_default_host();
610 phb->fwnode = irq_domain_alloc_named_id_fwnode("pSeries-MSI",
615 phb->dev_domain = irq_domain_create_hierarchy(parent, 0, count,
617 &pseries_irq_domain_ops, phb);
618 if (!phb->dev_domain) {
619 pr_err("PCI: failed to create IRQ domain bridge %pOF (domain %d)\n",
620 phb->dn, phb->global_number);
621 irq_domain_free_fwnode(phb->fwnode);
625 phb->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(phb->dn),
626 &pseries_msi_domain_info,
628 if (!phb->msi_domain) {
629 pr_err("PCI: failed to create MSI IRQ domain bridge %pOF (domain %d)\n",
630 phb->dn, phb->global_number);
631 irq_domain_free_fwnode(phb->fwnode);
632 irq_domain_remove(phb->dev_domain);
639 int pseries_msi_allocate_domains(struct pci_controller *phb)
643 if (!__find_pe_total_msi(phb->dn, &count)) {
644 pr_err("PCI: failed to find MSIs for bridge %pOF (domain %d)\n",
645 phb->dn, phb->global_number);
649 return __pseries_msi_allocate_domains(phb, count);
652 void pseries_msi_free_domains(struct pci_controller *phb)
655 irq_domain_remove(phb->msi_domain);
657 irq_domain_remove(phb->dev_domain);
659 irq_domain_free_fwnode(phb->fwnode);
662 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
664 /* No LSI -> leave MSIs (if any) configured */
666 dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
670 /* No MSI -> MSIs can't have been assigned by fw, leave LSI */
671 if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
672 dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
676 dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
677 rtas_disable_msi(pdev);
680 static int rtas_msi_init(void)
682 query_token = rtas_function_token(RTAS_FN_IBM_QUERY_INTERRUPT_SOURCE_NUMBER);
683 change_token = rtas_function_token(RTAS_FN_IBM_CHANGE_MSI);
685 if ((query_token == RTAS_UNKNOWN_SERVICE) ||
686 (change_token == RTAS_UNKNOWN_SERVICE)) {
687 pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
691 pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
693 WARN_ON(ppc_md.pci_irq_fixup);
694 ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
698 machine_arch_initcall(pseries, rtas_msi_init);